Comparing Uncertainty Estimation Methods in Deep Neural Networks

Official URL: https://risc01.sabanciuniv.edu/record=b3400662

Convolutional Neural Networks (CNNs) is one of the mainstream paradigms in most computer vision tasks. Accurately quantifying the uncertainty in CNN’s predictions is crucial as they are being used in various applications, including safety- critical domains such as medical image classification and autonomous driving. Yet, uncertainty prediction remains a challenge. Softmax probabilities are often used to model uncertainty with no solid support. Recent studies have tackled this challenge using three distinct methodologies, namely: Monte Carlo Dropout, Deep Ensembles, and Evidential Deep Learning (EDL). Although this thesis primarily focuses on EDL, the most up-to-date and computationally efficient among these approaches, each of these methods performance in uncertainty estimation along with their predictive capabilities are compared using CIFAR-10 and CelebA datasets in this work. Finally, leveraging the EDL method on the CelebA dataset, a novel approach is presented to automatically detect mislabeled samples within the dataset.